Vision-Language-Action (VLA) models have become a powerful paradigm for robotic manipulation and show strong promise for real-world deployment, making adversarial robustness critical for reliable robot operation. Existing adversarial attacks on VLA-controlled robots mainly optimize prediction-level discrepancies or rollout-level deviations to disrupt manipulation. However, they rarely account for the fact that VLA-predicted commands are not executed directly, but must be realized through downstream control and robot embodiment. In this paper, we propose ExecGap, an adversarial patch attack framework that exploits this command-execution gap to expose embodiment-specific execution vulnerabilities. Specifically, ExecGap biases VLA-predicted task-space commands toward low-executability directions, which are estimated from kinematic inefficiency and dynamic weak actuation. This induces execution degradation, where the predicted command becomes difficult to realize effectively, thereby impairing task completion. Extensive simulation experiments across multiple VLA policies show that ExecGap induces larger command-execution mismatch and weaker realized execution than existing patch attacks. Real-robot experiments further validate that the deployed patch can induce execution degradation on physical platforms.
Figures & tables
Figure 1: Command-execution gap in VLA-controlled manipulation. A VLA model predicts a task-space command ut from visual observation and language instruction. During embodied execution, constraints from control, kinematics, and actuation can reshape or attenuate the command, causing command-execution mismatch between the task-space command and the realized motion.
Figure 2: Overview of ExecGap. ExecGap first estimates low-executability directions from kinematic inefficiency and dynamic weak actuation. It then optimizes a visual patch to bias the VLA-predicted task-space command toward these directions, enlarging command-execution mismatch and inducing execution degradation.
Attack
Suite
TSR(%) ↓
CEDE( ∘ ) ↑
NCEE ↑
CEG ↓
MRE ↓
CRE ↓
Suite
TSR(%) ↓
CEDE( ∘ ) ↑
NCEE ↑
CEG ↓
MRE ↓
CRE ↓
Clean
Spatial
86.4
14.78
0.9836
0.0165
0.0085
1.0148
Goal
50.2
18.30
0.9767
0.0238
0.0045
1.5245
TMA
0.0
21.68
0.9824
0.0176
0.0026
0.5620
0.0
38.13
0.9677
0.0328
0.0004
0.2095
UADA
0.0
19.32
0.9839
0.0161
0.0025
0.5520
0.0
34.13
0.9692
0.0453
0.0006
0.3016
UMA
0.0
20.23
0.9842
0.0158
0.0022
0.4909
0.0
33.53
0.9658
0.0378
0.0007
0.3537
Ours
0.0
33.72
0.9896
0.0092
0.0020
0.4329
0.0
39.84
0.9930
0.0319
0.0004
0.2144
Clean
Object
76.4
14.83
0.9742
0.0264
0.0067
1.0760
Long
40.8
19.21
0.9770
0.0239
0.0043
1.5480
Table 1: Main comparison on OpenVLA across four LIBERO suites. The table reports task success rate and command-execution metrics. Best and second-best results among attack methods are marked in bold and underlined , respectively.
Figure 3: Qualitative rollout snapshots on LIBERO with OpenVLA. Representative snapshots from OpenVLA-controlled rollouts are shown under clean, UADA, and ExecGap settings.
Attack
Suite
OpenVLA-OFT
π0
Suite
OpenVLA-OFT
π0
TSR(%) ↓
CEDE( ∘ ) ↑
CRE ↓
TSR(%) ↓
CEDE( ∘ ) ↑
CRE ↓
TSR(%) ↓
CEDE( ∘ ) ↑
CRE ↓
TSR(%) ↓
CEDE( ∘ ) ↑
CRE ↓
Clean
Spatial
96.2
11.54
1.0172
94.4
11.87
1.0153
Goal
98.6
19.37
1.2181
94.4
17.93
0.8346
Rand.
88.6
13.65
1.1168
88.2
12.63
0.9539
94.6
19.62
1.2328
90.0
17.37
0.8682
TMA
61.6
25.29
0.6687
68.8
27.31
0.8153
72.0
21.47
0.8474
64.4
24.58
0.7759
UADA
66.8
25.78
0.7351
65.2
28.33
0.7860
70.0
22.93
0.7947
69.4
25.38
0.7675
UMA
63.2
24.28
0.6527
65.4
29.36
0.7984
73.8
22.46
0.7473
75.6
24.28
0.7234
Table 2: Comparison on OpenVLA-OFT and π0 across four LIBERO suites. The table reports task success rate and command-execution metrics. The best result among attack methods for each policy, suite, and metric is marked in bold .
Table 6
Figure 4: Real-robot rollout snapshots. Representative snapshots from OpenVLA-controlled executions on Kinova Jaco2 and ROKAE xMate ER7 Pro are shown under clean and ExecGap settings.
Adversarial attacks on vision-language-action (VLA) models study how crafted inputs, poisoned data, or physical cues can disrupt robot decisions and safety.
Vision-Language-Action (VLA) models have shown strong capabilities in controlling robots across diverse manipulation tasks. However, their adversarial robustness remains largely underexplored, and exploiting this weakness can lead to physical-world harm. Existing attacks on VLA models often rely on pixel-space perturbations or white-box access, resulting in noticeable artifacts and limited deployability in real-world robotic systems. In this work, we propose DURA, a diffusion-based unrestricted robotic attack that generates visually natural adversarial patches for VLA models. DURA supports both white-box and black-box attack settings, where the black-box setting requires only the predicted actions of the victim model. By optimizing along the latent trajectory of a pretrained diffusion model, DURA generates visually natural patches while steering the robot toward attacker-specified target actions. Extensive experiments in both simulation and the real physical world show that DURA consistently outperforms existing methods. Our findings expose a safety risk for physically deployed VLA models and call for stronger defenses.
Jiahui Han, Yuhui Yao, Xin Wang +6
1Xi’an Jiaotong University · 2Shanghai AI Laboratory · University of Science and Technology of China
Vision-language-action (VLA) models are gaining attention in robotics, yet their robustness to adversarial attacks remains largely unexplored. Existing work shows that adversarial patches can mislead VLA-based robots but assumes full access to the entire execution trajectory, an unrealistic requirement in practice. We address this limitation by formulating a partially observable threat model, where the adversary can exploit only a short prefix of the trajectory to generate a fixed patch applied to all subsequent frames. Under this setting, we propose a two-phase framework. First, we localize the patch using the model's attention maps to identify visually critical regions that correspond to the full instruction. Then, we optimize the patch to disrupt the semantic grounding of target objects and increase the curvature of action trajectories, thereby compounding failures in both perception and control. Extensive experiments in simulation and real-world robotic environments show that our method sustains adversarial effects under partial observability, inducing long-horizon disruptions and significantly reducing task success rates.
Xiaofei Wang, Mingliang Han, Tianyu Hao +3
Department of Automation, University of Science and Technology of China, Hefei, China · SmartMore Corporation · Cyberspace Institute of Advanced Technology, Guangzhou University, Guangzhou, China +2
Recently, Vision-Language-Action (VLA) models have revolutionized robotic manipulation by seamlessly integrating visual perception, language understanding, and action generation in an end-to-end learning framework. However, since these models are designed to interact directly with the physical world and humans, their security is critical, and even small vulnerabilities can lead to catastrophic failures. In this work, we propose the Universal Adversarial Object, a sphere with optimized surface texture that significantly degrades task success rates when placed within the robot's field of view. Specifically, our approach introduces a multi-level attack framework that jointly disrupts trajectory planning, task execution, and action control. We validate our method in both simulated and real-world robotic settings. Experimental results demonstrate that the adversarial object reduces the average task success rates by 31.2%-39.9% for two representative VLA models (Pi0 and RDT), with success rates dropping to near zero in complex scenarios. Index Terms--Vision-Language-Action models, adversarial attack, robotic security, universal adversarial object
Songhua Yang, Ziyu Liu, Yuanwei Liu +5
School of Computer Science, Wuhan University. · School of Robotics, Wuhan University. · School of Power and Mechanical Engineering, Wuhan University.